Post on 25-Oct-2020
transcript
Posted Wage Rigidity
Jonathon Hazell Mazi Kazemi Bledi Taska∗
February 2018
Abstract
This paper documents three new facts about posted wage rigidity, using a comprehensive and newly
available dataset from the United States. First, posted wages change infrequently. Wages for the typical
job remain unchanged for 20 quarters. Second, posted wages are especially unlikely to fall for a given
job, implying downwards rigidity in the posted wage. Third, posted wages are nearly acyclical for the
typical job. We derive sufficient statistics in a class of labour search models, and calibrate them with our
estimates. In the calibrated model, the estimated wage rigidity generates large fluctuations in unemploy-
ment over the business cycle. In future work, we plan to introduce similar datasets for the Euro Area, to
apply our methods to Euro Area wage setting.
∗Hazell: Massachusetts Institute of Technology, Department of Economics (email: jhazell@mit.edu); Kazemi: MassachusettsInstitute of Technology, Sloan School of Management; Bledi Taska: Burning Glass Technologies. We thank for their commentsDaron Acemoglu, Martin Beraja, Sydnee Caldwell, Chen Lian, Emi Nakamura, Jonathan Parker, Jon Steinsson, Olivier Wang andIvan Werning. The views here do not necessarily reflect those of Burning Glass Technologies.
1
1 Introduction
Why does unemployment rise during recessions? A leading hypothesis is wage rigidity. At the onset of
the typical recession, in both the US and within the Euro Area, hiring falls sharply. Unemployment subse-
quently rises. A leading explanation (Hall, 2005; Hall and Milgrom, 2008; Hagedorn and Manovskii, 2008;
Gertler and Trigari, 2009) for hiring fluctuations is rigidity in the wage for new hires, so that wages vary
little with the business cycle. The cost of hiring then remains high during recessions, even as labour de-
mand falls. Firms respond by hiring fewer workers, leading to a rise in unemployment. According to this
theory, the relevant quantity is the wage for new hires, because it measures the marginal cost of adding
a new worker. The wage for incumbent workers, who have already been hired, is less important. In New
Keynesian models of the business cycle (Christiano et al., 2005; Smets and Wouters, 2003, 2007), wage
rigidity also plays an important role in amplifying unemployment fluctuations.
Despite its importance in theory, we understand little about the empirical behaviour of the wage for
new hires1. This paper takes up the challenge of measuring the wage for new hires, and traces out the
implications for unemployment fluctuations. We present three new facts, which suggest substantial wage
rigidity. We then map the estimated rigidity to a model, and show that it rationalises large fluctuations in
unemployment over the business cycle.
We start by introducing a new and comprehensive dataset of posted wages. We use a proprietary
dataset of online job vacancy postings, provided by Burning Glass Technologies, a labour market analytics
firm. The dataset has numerous advantages compared with existing survey or administrative datasets.
The data contains posted salaries, with both posted hours worked and bonus or over time pay where
applicable. By contrast, administrative data often does not record hours worked, and survey data tends to
have noisy measures of both pay and hours—hampering measurement of the cost of labour. The posted
wage dataset is larger than existing surveys, covering roughly 10% of vacancies posted either online or
offline in the United States since 2010. We observe the job title and establishment2 of the job posting. We
can then study multiple wage postings for the same job, allowing us to observe the rigidity of wages at the
job level.
We present three new stylised facts. First, posted wages change infrequently. Wages for the typical
job remain unchanged for 20 quarters. Second, posted wages are especially unlikely to fall within a given
job, implying downwards rigidity in the posted wage. Third, posted wages are nearly acyclical for the
typical job. Overall, these three new facts suggest substantial rigidity in the wage for new hires, at the job
level. To our knowledge, the findings have never previously been documented for posted wages. Previous
work on wage rigidity has been largely limited to incumbent workers, which is not helpful for testing
the relevant theory. By contrast, our estimates of posted wage rigidity directly pertain to the relevant
quantity, the wage for newly hired workers. We are able to uncover the new facts because of the unique
feature of dataset—whereby we can observe multiple wage postings for the same job, and so understand
the strength of rigidity at the job level. The previous literature tends to find that the wage for new hires is
procyclical. We show that cyclical changes in job composition, and imprecise estimates, makes previous
findings of procyclicality hard to interpret.
We then feed our estimates into a model, to understand the implications for unemployment fluctua-
1There is a large literature on wage rigidity for incumbent workers, including Card and Hyslop (1997) for the United States,Le Bihan et al. (2012) for France, Bauer et al. (2007) for Germany and Devicienti et al. (2007) for Italy. However wage rigidity for newhires—the relevant concept according to benchmark theories—is less studied.
2Each physical location at which a firm employs workers is a separate establishment.
2
tions. We show that in a large class of labour search models, the cyclicality of posted wages is a key force
determining unemployment fluctuations. When we calibrate our model to the estimated posted wage
rigidity, fluctuations in hiring are large, in line with the US time series data. Thus posted wage rigidity is
an important contributor to the volatility of unemployment over the business cycle.
In the future, we plan to apply our methods to Euro Area wage setting. As previously discussed, our
dataset of online vacancy postings has unique advantages for measuring posted wage rigidity. Moreover,
we have developed a theoretical framework for relating estimated wage rigidity to unemployment fluctu-
ations. Through Burning Glass Technologies, we will soon gain access to similar vacancy posting data for
France, and potentially other Euro Area countries in the future. Hence we can use the methods developed
in this paper to assess whether posted wage rigidity can rationalise large fluctuations in unemployment
in the Euro Area.
Related literature. This paper relates to the literature on wage rigidity and unemployment fluctu-
ations. An influential paper by Shimer (2005) argues that benchmark models of labour search cannot
rationalise large unemployment fluctuations. Various papers—in particular by Hall (2005), Hall and Mil-
grom (2008), Hagedorn and Manovskii (2008) and Gertler and Trigari (2009)—argue that wage rigidity for
new hires can rationalise large fluctuations in hiring and unemployment. These papers use variants of
the canonical labour search model, the Diamond-Mortensen-Pissarides model, to study unemployment
fluctuations. A parallel literature has studied the effect of wage rigidity on unemployment in New Key-
nesian models with both sticky prices and wages—key contributions include Smets and Wouters (2003),
Christiano et al. (2005) and Smets and Wouters (2007). Christiano et al. (2016) unite the two modelling
frameworks, and show that labour search with rigid wages, and sticky prices, can explain many key busi-
ness cycle moments. Nevertheless, the importance of wage rigidity is contested. Pissarides (2009) argues
that estimated wage rigidity is not large enough to rationalise large hiring fluctuations. Chodorow-Reich
and Karabarbounis (2016) show that existing models imply a procyclical opportunity cost of employment,
which nullifies endogenous wage rigidity.
This paper also speaks to the literature on downwards wage rigidity. Many papers argue that down-
wards wage rigidity implies asymmetric effects of labour demand on unemployment. Two leading exam-
ples are Chodorow-Reich and Wieland (2016) and Dupraz et al. (2016). In these papers, the relevant wage
is the wage for new hires. However, outside our paper, there is minimal evidence of downwards rigidity
for new hires. Our paper provides new evidence supporting the key assumption in this literature.
Finally, our work belongs to the literature estimating the cyclicality in the wage for new hires. Pre-
vious work tends to find strongly procyclical wages for new hires, both in the US and in the Euro Area.
Examples in the US include Bils (1985), Shin (1994), Haefke et al. (2013) and Hagedorn and Manovskii
(2013). Examples in the Euro Area include Peng and Siebert (2008), Martins et al. (2012) and Carneiro
et al. (2012). Equally, when measuring the cyclical of the hiring wage for workers entering from unem-
ployment, Gertler et al. (2016) estimate a nearly acyclical wage in US data. We argue that imprecision and
cyclical change in job composition makes these estimates hard to interpret.
2 Dataset
Our main resource is a proprietary dataset of online vacancy postings, provided by Burning Glass Tech-
nologies (BGT). The coverage is 2010-2016. BGT uses machine learning algorithms to extract vacancy
posting data from online job boards, and company websites. Independent work (Carnevale et al., 2014)
3
finds that the BGT’s algorithms correctly classify a high share of job postings. The job posting data con-
tains posted salaries, which includes a measure of hours worked. Posted salaries are classified as hourly,
weekly, monthly or annual. The salary includes bonus or overtime pay where applicable. These features
of the wage data are an advantage compared with existing datasets. Administrative data often does not
contain measures of hours worked, which is necessary for uncovering the marginal cost of labour. Survey
data tends to have noisy measures of both hours worked and salary payments, especially when includ-
ing bonus and overtime pay. The posting data reports both establishment and job title. Each physical
location at which a firm employs workers is an establishment. An establishment is therefore a location
identifier. Job titles are standardised using BGT’s algorithm.
An example clarifies the granularity of the dataset. Consider a large firm, such as Costa Coffee, which
has many physical locations across the United States, and hires for many positions, such as baristas or
managers. For each vacancy posting by Costa, we can observe the establishment, i.e. the physical lo-
cation; the job title, e.g. barista or manager; the salary, inclusive of bonus if applicable; and the pay
frequency, e.g. hourly or annual.
The dataset of posted wages is also large, covering roughly 10% of vacancies posted either online or
offline in the US (Carnevale et al., 2014). A key advantage of this dataset is that we see how posted wages
for the same job vary over time. We can then study rigidity in job-level posted wages. The dataset also
contains industry and occupation information about the vacancy posting. The industry information is at
the 2- 4- and 6-digit NAICS code level. Occupation information is at the 2- 4- or 6-digit SOC code level.
2.1 Validating the Posted Wage Data
The posted wage data matches variation in actual US wages. We compare wages by occupation. We
study occupation at the six-digit SOC level3. We take the median posted wage within each occupation
for Burning Glass for 2010-2016; and the median hourly wage within occupation from the 2014-2016
Occupational Employment Statistics (OES), the establishment-level survey of occupational wages in the
US. We regress OES wages on Burning Glass wages, by occupation. The results are in Appendix Table 4
and Appendix Figure 6. Wages by occupation in Burning Glass closely match the OES.
We also compare wages by region. We study regions at the core-based statistical area4 (CBSA) level.
We take the median posted wage within each CBSA for Burning Glass for 2010-2016; and compare to
the 2010-2016 Quarterly Census of Employment and Wages (QCEW), the regional census of wages in the
United States. We regress QCEW wages on Burning Glass wages, by CBSA. The results are in Appendix
Table 5 and Appendix Figure 7. Wages by CBSA in Burning Glass closely match the QCEW.
2.2 Representativeness
Next, we study the representativeness of our dataset. Appendix Figure 8 plots the relative share of Burning
Glass occupations, at the 2-digit SOC level, versus the 2014-2016 Occupational Employment Statistics.
Burning Glass overweights transportation, healthcare, computation, and finance; and underweights
construction, education, and food preparation. Where important for robustness, we reweight to target
the US occupation distribution, to deal with issues of data representation.
3These occupations are granular, at the level of, for example, a high school Spanish teacher.4A CBSA is an urban area, either a micropolitan or metropolitan statistical area. It is defined by commuting ties, to accurately
capture the local labour market.
4
Table 1: Summary Statistics, Data Differenced by Job
Min Max AverageTotal Number 1319756Occupation Coverage .97Postings Per Job 2 25 2.76Jobs Per Occupation 1 192471 1144Jobs Per CBSA 1 24175 760
Notes: Occupation is by 6-digit SOC code. Occupation share is the total share of 6 digit SOC occupations,by employment in the 2014-16 OES, which are represented in the Burning Glass data. Burning Glass datais 2010-2016. A job is an establishment by job title by pay frequency by salary type unit. Posted wages areaveraged within each job-quarter.
2.3 Regional and occupational data
In many specifications, we will use regional business cycle variation. We will use regional unemployment
from the Local Area Unemployment Statistics (LAUS), regional employment from the Quarterly Census
of Employment and Wages (QCEW), regional industry-by-employment shares from the County Business
Patterns (CBP), and national industry employment from the Current Employment Statistics (CES).
Finally, we use occupational data from the 2014-2016 Occupational Employment Statistics (OES).
3 Empirical Results
Fact 1: Posted Wages Change Infrequently
This section introduces the first new fact about posted wage rigidity. For a given job, posted wages change
infrequently—rather, they remain constant over many vacancy postings, and for a long period of time.
Though this fact has previously been documented for incumbent workers, this paper is the first to docu-
ment infrequent wage changes in posted wages.
Firstly, we discuss measurement details. We define a job as a job-title by establishment by pay category
unit. So, using the example of a Costa Coffee barista, a job is a barista at a given physical location of Costa,
with an hourly wage. We aim to study multiple posted wages for the same job, and therefore restrict to jobs
with multiple wage postings. We take the mean posted wage within each job-quarter. This step sweeps
out high frequency variation in wages, due to, for example, multiple vacancy postings by the same firm
within the same month. After these steps, there are roughly 1.4 million postings remaining. Table 1 details
summary statistics for this restricted sample.
Figure 1 presents the first fact. Within a given job, posted wages change infrequently. In the figure,
the x-axis is the growth in the posted wage between two consecutive job postings by the same job. The
graph shows the entire distribution of posted wage growth in the sample. The y-axis is the frequency of
observations. The posted wage growth distribution is truncated at the 5th and 95th percentile. The graph
shows that for most jobs, the posted wage growth is at or near zero.
This fact is new. Infrequent wage changes have previously been documented for incumbent work-
ers—but never for repeated wage postings for the same job. This fact is the first piece of evidence indi-
cating substantial rigidity at the job level. We can document this new stylised fact because of the unique
feature of our dataset, whereby we observe multiple vacancy postings for the same job.
5
Figure 1: Posted Wages Change Infrequently
8060
4020
Perc
ent
-20 -10 0 10 20Posted Wage Growth
Notes: this graph measures the distribution of the growth in posted wages between two consecutive post-ings by the the same job. As before, a job is an establishment by job-title by salary type by pay frequencyunit. The salary growth distribution is truncated at the 10th and 90th percentiles.
An alternative way to see the first key fact is through calculating summary statistics, presented in Table
2 . The typical wage posting spell is long. Posted wages within a job remain unchanged for long periods
of time, spanning multiple vacancies.
For each job, we calculate the probability that the posted wage changes, after a new vacancy. We then
take the median across jobs, to arrive at the overall probability that the wage changes on a new vacancy.
We then calculate the median number of vacancy postings for which the posted wage is unchanged. In
particular, we calculate the median implied duration of a posted wage spell, in terms of the number of
vacancies, by inverting the probability5 of posted wage change on a new vacancy. Finally, we report the
median duration of a posted wage spell in terms of the number of quarters.
The key takeaway is that posted wages change infrequently within a given job. The probability that a
wage changes on a new vacancy is 0.7. Meanwhile, posted wages for the typical job are unchanged for 20
quarters. The result is robust to multiple ways of calculating the main summary statistics6.
Fact 2: Downwards Rigidity in Posted Wages
We now establish a second fact about posted wage rigidity at the job level. This section documents sub-
stantial downwards posted wage rigidity at the job level. The job level rigidity in posted wages, docu-
mented in the previous section, has asymmetric effects—and prevents firms from cutting posted wages.
5We use the implied duration formula d = −1/ log(1 − f) where d is the duration, and f is the probability that a posted wagechanges.
6E.g. by weighting the medians to target the US occupation distribution from the Occupational Employment Statistics.
6
Table 2: Posted Wage Setting Statistics
Duration of Median Posted Wage Spell, in Quarters 19.5Number of Vacancies in Median Posted Wage Spell 13Probability of Posted Wage Change for Median Job .07Number of observations 1319756
Notes: a posted wage spell is the number of vacancy postings for a which a posted wage remains un-changed. The median implied duration inverts the median probability of posted wage change, and isgiven by the formula d = − 1
log(1−f) where f is the frequency of posted wage change. Posted wages areaveraged within each job-quarter.
Within a given job, posted wages are especially unlikely to fall. Figure 2 presents the main fact. From
the graph, for most jobs the posted wage is more likely to rise than to fall.
To construct the graph, we take the distribution of posted wages as in Figure 1. We then exclude posted
wage changes of zero, to leave only non-zero posted wage changes. Finally, we truncate the posted wage
growth distribution at ±10%. In the graph, posted wages are more likely to increase than to decrease.
Moreover, the probability of a small increase is discontinuously higher than the probability of a small
decrease in posted wages.
Again, this fact is new. Downwards rigidity was previously shown for incumbent workers (e.g. Le Bi-
han et al., 2012) but never for new wages. This second fact also supports the overall narrative of the
paper—that posted wage rigidity is large at the job level. Again, we can document this new stylised fact
because of the unique feature of our dataset, where by we observe multiple vacancy postings for the same
job.
We next devise a test to underscore the importance of downwards wage rigidity. If downwards rigidity
is important, then the probability of lowering posted wages should be low, and insensitive to labour de-
mand. Since firms are constrained, they should rarely lower wages, regardless of labour demand. Mean-
while, the probability of raising posted wages should be more sensitive to labour demand. Firms are
unconstrained upwards, and can raise wages in response to positive labour demand shocks. We find
support for both these predictions.
We start by constructing regional labour demand measures. We construct regional measures of labour
demand following Bartik (1991). For each CBSA, we calculate the employment share by industry, at the
2-digit NAICS level for 2007. We then weight national 2-digit NAICS industry employment growth for
2010-2016 by the regional weights7. The result is a CBSA-specific measure of the cumulative increase in
labour demand over 2010-2016. The measure captures labour demand, and not labour supply, provided
that regional labour supply shocks over 2010-2016 are orthogonal to the 2007 industry shares.
The probability of a wage rise increases with labour demand, consistent with firms being less con-
strained when raising wages. Figure 3 presents this fact. On the x-axis is the regional labour demand
shock, in percentiles. On the y-axis is the median probability of a posted wage rise, in each CBSA. The
graph is a binned scatterplot, with 5% bins of the labour demand shock, and a non-parametric regression
estimate. The probability of a wage rise is high in high labour demand regions—which holds if firms are
unconstrained when trying to increase posted wages for a given job.
By contrast, the probability of a wage fall is low and does not change with labour demand, consistent
7The 2007 regional weights are from the County Business Patterns. The 2010-2016 industry employment growth is from theCurrent Employment Statistics.
7
Figure 2: Posted Wages Are Rigid Downwards
0.0
2.0
4.0
6.0
8.1
Estim
ated
Ker
nel D
ensi
ty
-10 -5 0 5 10Posted Wage Growth, Excluding Zero Growth
Notes: this graph measures the distribution of the growth in posted wages between two consecutive post-ings by the the same job, excluding zero growth observations. As before, a job is an establishment byjob-title by salary type by pay frequency unit. The salary growth distribution is truncated at ±10%. Kerneldensity estimation uses an Epanechnikov kernel with a bandwidth of 0.65.
with downwards rigidity. The median posted wage change is zero in all CBSAs, regardless of the labour
demand in the CBSA. Thus the probability of a posted wage fall within the job, is low and insensitive to
labour demand—confirming the importance of downwards wage rigidity.
Overall, there is substantial rigidity in posted wages. Not only do they change infrequently at the job
level, but they are especially unlikely to fall.
Fact 3: Posted Wages are Nearly Acyclical
We already documented two new stylised facts: posted wages infrequently change for a given job, and are
especially unlikely to fall. We now show that wages vary little within a given job over the business cycle.
Cyclicality in posted wages captures firms’ incentives to hire for the same job at different stages of the
business cycle. Acyclical posted wages imply large fluctuations in hiring incentives. Therefore the rigidity
previously documented at the job level has important business cycle implications8.
To measure wage cyclicality, we build on the canonical regression of Bils (1985). Following this paper,
the literature estimates the regression
logwit = α+ βUt + controlsit + εit,
8As suggested by Caplin and Spulber (1987) and Golosov and Lucas Jr (2007), infrequent posted wage changes at the job levelmay not have important cyclical implications, if firms optimally time when they change wages. By contrast, our estimates suggestthat the job level rigidity has important cyclical implications.
8
Figure 3: Probability of Wage Rise Increases in Labour Demand
0.0
5.1
.15
.2M
edia
n P
roba
bilit
y of
Pos
ted
Wag
e R
ise
by C
BS
A
25 50 75 100Percentile of Labour Demand Shock
Notes: the probability of posted wage rise is the median by CBSA. Labour demand is a Bartik/shift-sharemeasure, calculated with 2007 2-digit NAICS employment shares by CBSA, and 2010-2016 national 2-digit NAICS employment growth. Labour demand shocks are collected in 5% bins. The conditional meanprobability of posted wage change is the mean probability within each bin, weighted by 2016 employmentfrom the QCEW. The nonparametric regression line is estimated by lowess, with a bandwidth of 0.8.
wherewit is the cyclical component of the wage for new hires, andUt is the cyclical component of national
unemployment. β, the semi-elasticity of wages with respect to unemployment, then measures cyclicality
in the wage for new hires.
We adapt the procedure to our setting in three ways. Firstly, we argued that wage variation within
a given job is the relevant variation for studying change in incentives to hire for that job over the cycle.
Wage variation between jobs may be less relevant. Therefore we only study within-job wage variation in
our regressions—which is only possible because of the unique feature of our dataset, whereby we can ob-
serve multiple vacancy posts for the same job. Secondly, there are limited national business cycles in the
United States over 2010-2016. We therefore harness extra variation from regional business cycles. Finally,
regional unemployment is noisily measured in the United States. We therefore project unemployment
onto a high quality administrative measure of employment.
Overall, our specification is a regional version of the canonical regression of Bils (1985), using only
within-job variation. The regression is
∆ logwijt = α+ controlsjt + β∆Ujt + εjt,
where wijt is the nominal posted wage in job i and CBSA j in quarter t. As before, a job is a job-title by
9
Table 3: Quarterly Posted Wage Cyclicality, Differenced By Job
Dependent Variable: Posted Wage Growth, by Job(1) (2) (3) (4) (5)
Independent Variable:Quarterly Unemployment Change -0.0856 -0.122 -0.146 0.0314 -0.221
(0.0917) (0.0843) (0.111) (0.0739) (0.166)Seasonal Dummies Y Y Y Y YDifference Length Dummies Y Y Y Y YTime Effects N Y N N NOES Weights N N Y N NCBSA Fixed Effects N N N Y NWinsorized N N N N YNumber of Differenced Observations 1211948 1211948 1204026 1209224 1237127
Notes: the dependent variable is percentage posted wage growth 100 × ∆ log (wijt) , for job i in CBSA jat quarter t, from the 2010-2016 Burning Glass data. Posted wages are averaged within each job-quarter.The independent variable is the change inUjt, the quarterly unemployment rate in CBSA j at time t, fromthe 2010-2016 LAUS. We projectUjt onto quarterly employment growth from the 2010-2016 QCEW. Postedwage growth is trimmed at the 1st and 99th percentile, except in column (5), in which they are Winsorizedat the 1st and 99th percentiles. In column (3), the OES weights reweight the Burning Glass data to matchthe 2014-2016 OES at the 6-digit SOC level. A job is an establishment by job title by pay frequency bysalary type unit. Standard errors are in parentheses, two-way clustered by CBSA and quarter. One, twoand three asterisks denote significance at the 10, 5 and 1 percent levels, respectively.
establishment observation. Ujt is unemployment in CBSA j and quarter t. ∆ logwijt is differenced by
job, and ∆Ujt is differenced by CBSA. We project ∆Ujt onto ∆ log(
Employmentjt), which is CBSA em-
ployment growth from the QCEW. By design, this regression uses our dataset to only focus on within job
variation. It uses regional business cycle variation. It deals with measurement error in unemployment, by
projecting onto regional employment growth. Then β measures posted wage cyclicality. By running the
regression in first differences, we also avoid issues with nonstationarity and a persistent error process.
We estimate acyclical posted wages. Table 3 presents the results. In our benchmark specification, after
a percentage point fall in quarterly regional unemployment, wages within the typical job growth by only
0.08%. Though posted wages are procyclical, the degree of procyclicality is small. This finding is robust
to numerous specifications, including adding CBSA specific trends or time effects, reweighting to target
the distribution of jobs in the US economy, different forms of dealing with outliers, using annual data, or
using different measures of the local labour market other than CBSAs. These additional robustness tests
are in Appendix Section B. Overall, the rigidity previously documented at the job level means that posted
wages are nearly acyclical within the job. Figure 4 displays the results from our benchmark regression in
a binned scatterplot.
Composition Bias, Precision and the Literature
We now relate our findings to the previous literature. The previous literature typically finds that wages
for new hires are procyclical. We argue that these estimates are biased upwards by changes in the cyclical
composition of jobs—and that we are able to overcome this bias due to the unique features of our dataset.
Moreover, our relatively precise estimates are easier to interpret than those of the preceding literature.
In this paper we are interested in how incentives to hire for a a given job change over the cycle. The
10
Figure 4: Posted Wages are Nearly Acyclical
-.2
-.1
0.1
.2P
oste
d W
age
Gro
wth
-.5 0 .5Unemployment Change
This graph is a binned scatterplot of the regression presented in Table 3, column (1).
incentives to hire for a job depend on how the labour cost of that job, i.e. the hiring wage, changes with
the cycle. To calculate this cyclicality, we must focus only on within-job variation—or conversely, we must
hold constant the cyclical composition in jobs.
By contrast, cyclical changes in the composition of job quality give a procyclical bias to estimates of
wage cyclicality. Suppose that higher wage jobs are created during booms than busts. This shift in job
composition over the cycle will cause wages to be higher during a boom than a bust, and so will generate
wage cyclicality. However, this wage variation will not capture changes in incentives to hire within a given
job over the cycle—but rather incentives to shift between jobs. This variation is less helpful for studying
hiring incentives at the job level.
Previous work attempts to deal with this issue by directly introducing compositional controls, using
survey data. However, these controls are relatively coarse, and may be insufficient to control for cyclical
variation. By contrast, a unique advantage of our dataset is that we observe repeat wage postings for the
same job. We can study within-job variation directly, and therefore fully control for cyclical changes in
job composition.
We find estimates of the new hire wage cyclicality that are much less cyclical than the preceding liter-
ature. Figure 5 compares our estimates with six leading papers that study the cyclicality of the new hire
wage for the United States. Appendix Table 8 reports the values from the literature.
In the graph, the blue dots are point estimates and the red bands are 95% confidence interals. A more
negative value indicates greater procyclicality. Our estimate is the least procyclical. Other papers control
for cyclical changes in the composition of jobs to a varying extent, and with varying success, using survey
data and coarse compositional controls. Thus a procyclical bias from job composition can explain the
difference between our results and those of the preceding literature. Indeed, when we re-run our bench-
mark specification without conditioning on within-job variation, we uncover the procyclical estimates
11
Figure 5: Comparison of New Hire Wage Cyclicality with the Literature
-8
-6
-4
-2
0
2
Unemployment Semielasticity of Wages for New Hires
Gertler et al (2016)
Hagedorn & Manovskii (2013)
Haefke et al (2013)
Barlevy (2001)
Bils (1985) Shin (1994)
Our benchmark estimate
Greater Wage
Cyclicality
from the preceding literature—confirming that composition bias generates the previous procyclical esti-
mates. The special features of our dataset, with multiple wage postings for the same job, allow us to draw
this comparison.
Our dataset offers a second benefit—substantially less noisy estimates. In figure 5, our estimates are
much more precise than the previous literature. Moreover our standard errors are clustered to account
for correlated residuals in the panel and cross-section. Therefore with our estimates, we can reject mean-
ingful wage cyclicality. The greater precision comes from three sources. Firstly, we gain extra precision
from regional business cycle variation. Secondly, by studying within-job variation, we eliminate residual
noise on wages. Thirdly, we have many more observations than the preceding literature.
Overall, we find that posted wages within a given job are nearly acyclical. The preceding literature
finds that the wage for new hires are procyclical. The discrepancy between our and the previous findings
is explained by the failure of the previous literature to account for cyclical changes in job composition.
4 Posted Wage Rigidity in a Model
We showed that there was substantial rigidity in posted wages at the job level, and that this rigidity led to
acyclical wages for a given job. We can now revisit the original motivation: can the posted wage rigidity
estimated in the data rationalise large fluctuations in hiring? We derive a model which nests a wide class
of labour search models. We show that in this class of models, the rigidity of the posted wage is a key force
which amplifies hiring fluctuations. Finally, we show that in a plausible calibration, our estimated wage
rigidity generates large hiring fluctuations. Thus rigid posted wages leads to large hiring fluctuations.
This is the first paper, to our knowledge, that estimates rigidity in the new hire wage, and finds that it can
rationalise large unemployment fluctuations.
12
4.1 Model Setup
The model is in discrete time, and nests a wide class of labour search models. Firms post vacancies,
unemployed workers search for vacancies, and with some probability firms and workers match to form
a job. The match lasts for an uncertain number of periods, and ends with fixed probability s in every
period. The match produces output yt+j in period t + j. Firms pay workers wt,t+j , for a wage in period
t + j and a match starting in period t. At the beginning of the match, firms pay a fixed cost of matching
Ht. In this model, yt is the output per worker from the match—and so is a measure of labour demand.
Let Vt be the value of an unfilled vacancy, and Jt be the value of a filled vacancy. The value of an
unfilled vacancy is given recursively by
Vt = −γ + q(θt)Jt + (1 − q(θt)βEtVt+1, (1)
where q(·) is the probability that a vacancy is filled, and γ is a vacancy posting cost. θt is market tightness,
defined by θt ≡ vtut, where vt is the total number of vacancies, and ut is the total number of unemployed
workers search for jobs. q(·) is decreasing in θ. When the labour market is tight, with many vacancies
relative to unemployed workers, the probability of filling a given vacancy is low.
We assume a free entry condition, that the value of a vacancy is always zero in equilibrium, so that
Vt+j = 0 for all j. Then equation 1 simplifies to
Jt =γ
q(θt).
Under free entry, the value of a job is given by the cost of posting a vacancy for that job, γ, scaled by the
probability of filling the vacancy.
The value of a job is also the present value of the output from that job. We have
Jt =
∞∑j=0
[β(1 − s)]j
(yt+j − wt,t+j) −Ht.
The value of a job to a firm is the present value of the match output, after deducting wage payments, and
the initial matching cost—and discounting to account for the probability that the match ends.
This framework nests a wide class of models. It captures the benchmark competitive search model
(Moen, 1997). It also captures many common variants of the Diamond-Mortensen-Pissarides, including
the models of Shimer (2005); Hall (2005); Hall and Milgrom (2008); Hagedorn and Manovskii (2008); Pis-
sarides (2009); Christiano et al. (2016) and Chodorow-Reich and Karabarbounis (2016). The wage setting
process is consistent with either wage posting or wage bargaining.
4.2 A Simple Formula
We now derive a simple formula to understand whether our estimated wage rigidity can rationalise large
fluctuations in hiring, in response to changes in labour demand. Our formula maps from wage rigidity
to fluctuations in tightness while nesting a wide class of models. It is therefore robust to the underlying
details of the wage setting mechanism, and so allows us to map from wage rigidity to hiring while making
relatively few assumptions.
Purely for ease of exposition, we make two assumptions:
13
1. The cyclical component of yt is a random walk.
2. Wages are acyclical within the match.
Both assumptions are made purely to present a more simple formula. Equally, they both have sup-
port.Shimer (2005) and Ljungqvist and Sargent (2017) report that random walk assumption is a good
approximation in this setting. Numerous papers have found that wages are sticky within a given match
(e.g. Gertler and Trigari, 2009). Now, we derive a simple formula.
Proposition 1. The elasticity of market tightness with respect to labour demand is
d log θtd log yt
=1
α
Wage Rigidity Channel︷ ︸︸ ︷(1 − dwp
t
dyt
)yt
yt −Kt︸ ︷︷ ︸Profit Share Channel
.
Here, wpt is the posted wage. α is the elasticity of the probability of vacancy-filling with respect to
market tightness θ. Kt ≡ wpt + (1 − β(1 − s))Ht is the average cost of labour of a match starting at t,
inclusive of the initial fixed costs of matching.
This formula explains what determines the cyclicality of market tightness. Market tightness, in turn,
governs hiring and unemployment—when tightness is high, hiring is large relative to unemployment,
and unemployment falls rapidly. Therefore this formula explains what factors cause hiring and unem-
ployment to be sensitive to labour demand, in a wide class of labour search models.
The first factor is wage rigidity. When wages for new hires are rigid, they remain high during a down-
turn, and low during a boom. Meanwhile, labour demand is low during a downturn and high during a
boom. Hence the marginal incentive to hire is much larger in a boom than a downturn, leading to large
fluctuations in hiring—and hence unemployment—over the business cycle. Many papers, including Hall
(2005), Hall and Milgrom (2008) and Gertler and Trigari (2009), use this insight to rationalise large fluctu-
ations in unemployment. Importantly, conditional on a given posted wage, wages for incumbent workers
have no further implications for market tightness and hiring. As we previously argued, the relevant quan-
tity is the wage for new hires, or the posted wage.
The second factor is a small profit share. When profits are small, a given change in output per worker
has a larger proportional impact on profits. Thus profits respond more elastically, when they are small
on average. When profits are sensitive to business cycle fluctuations, hiring is also sensitive—since high
profits lead to increases in hiring, and low profits reduce hiring. Papers such as Hagedorn and Manovskii
(2008) and Pissarides (2009) use this insight to rationalise hiring fluctuations.
Overall, in wide class of labour search models, we have characterised the sensitive of market tightness
to labour demand, in terms of two sufficient statistics, dwpt
dytand yt
yt−Kt.
4.3 Calibration
We now have, for a large class of models, a simple formula mapping from posted wage rigidity to fluc-
tuations in market tightness—which captures movements in hiring. We now calibrate the model to our
estimated posted wage rigidity, and ask if the resultant fluctuations in hiring are large. We find that they
are. Therefore our estimated posted wage rigidity can rationalise large hiring fluctuations.
14
We calibrate with consensus values, where possible. We choose α = 0.5, matching the values of
Petrongolo and Pissarides (2001) and Sahin et al. (2014). We set Kt = 0.7, to match the average labour
share in the US economy.
Finally, we calibrate a value of dwpt
dytusing our estimates. For simplicity, we set dwp
t
dyt= 0, to match our
estimate of an acyclical posted wage. From our model, the implied elasticity of tightness with respect to
labour demand isd log θtd log yt
= 6.6.
Meanwhile, in the US time series data, the estimated value from Pissarides (2009) is
d log θtd log yt
= 7.56.
Thus our estimated wage rigidity generates large fluctuations in market tightness, in line with the time
series data. Overall, estimated wage rigidity then generates substantial and quantitatively realistic fluctu-
ations in hiring.
5 Conclusion
We introduced a new dataset of posted wages from online vacancies. This dataset has significant advan-
tages, relative to existing survey datasets. We are able to study multiple wage postings for the same job,
and control for the composition of jobs over the business cycle.
We use this dataset to answer a fundamental question in the macroeconomics of labour markets: can
the rigidity of wages for new hires rationalise large unemployment fluctuations? We document three
new stylised facts. Firstly, posted wages are rigid. For the typical job, the posted wage is unchanged for
20 quarter. Secondly, posted wages are especially rigid downwards. Thirdly, posted wages are acyclical
for the typical job. We show that previous estimates of a procyclical wage for new hires are difficult to
interpret, due to imprecision and a bias from cyclical changes in job composition.
We map our estimated wage rigidity into a model. The model transparently maps from rigidity in
the wage of new hires, to hiring fluctuations, in a large class of models. The estimated wage rigidity
rationalises large fluctuations in hiring and unemployment, in line with the data.
In future work, we hope to introduce a similar dataset of posted wages for the Euro Area. We will then
use the methods developed in this paper to assess whether wage rigidity in the Euro Area contributes
to large hiring fluctuations. A natural next step is to understand which policies can mitigate volatility in
unemployment over the business cycle.
15
References
Timothy J Bartik. Who benefits from state and local economic development policies? 1991.
Thomas Bauer, Holger Bonin, Lorenz Goette, and Uwe Sunde. Real and nominal wage rigidities and the
rate of inflation: Evidence from west german micro data. The Economic Journal, 117(524), 2007.
Mark J Bils. Real wages over the business cycle: evidence from panel data. Journal of Political economy,
93(4):666–689, 1985.
Andrew S Caplin and Daniel F Spulber. Menu costs and the neutrality of money. The Quarterly Journal of
Economics, 102(4):703–725, 1987.
David Card and Dean Hyslop. Does inflation" grease the wheels of the labor market"? In Reducing
inflation: Motivation and strategy, pages 71–122. University of Chicago Press, 1997.
Anabela Carneiro, Paulo Guimarães, and Pedro Portugal. Real wages and the business cycle: Accounting
for worker, firm, and job title heterogeneity. American Economic Journal: Macroeconomics, 4(2):133–52,
2012.
Anthony P Carnevale, Tamara Jayasundera, and Dmitri Repnikov. Understanding online job ads data.
Georgetown University, Center on Education and the Workforce, Technical Report (April), 2014.
Gabriel Chodorow-Reich and Loukas Karabarbounis. The cyclicality of the opportunity cost of employ-
ment. Journal of Political Economy, 124(6):1563–1618, 2016.
Gabriel Chodorow-Reich and Johannes Wieland. Secular labor reallocation and business cycles. Techni-
cal report, National Bureau of Economic Research, 2016.
Lawrence J Christiano, Martin Eichenbaum, and Charles L Evans. Nominal rigidities and the dynamic
effects of a shock to monetary policy. Journal of political Economy, 113(1):1–45, 2005.
Lawrence J Christiano, Martin S Eichenbaum, and Mathias Trabandt. Unemployment and business cy-
cles. Econometrica, 84(4):1523–1569, 2016.
Francesco Devicienti, Agata Maida, and Paolo Sestito. Downward wage rigidity in italy: Micro-based
measures and implications. The Economic Journal, 117(524), 2007.
Stephane Dupraz, Emi Nakamura, and Jon Steinsson. A plucking model of business cycles. Technical
report, mimeo, 2016.
Mark Gertler and Antonella Trigari. Unemployment fluctuations with staggered nash wage bargaining.
Journal of political Economy, 117(1):38–86, 2009.
Mark Gertler, Christopher Huckfeldt, and Antonella Trigari. Unemployment fluctuations, match quality,
and the wage cyclicality of new hires. Technical report, National Bureau of Economic Research, 2016.
Mikhail Golosov and Robert E Lucas Jr. Menu costs and phillips curves. Journal of Political Economy, 115
(2):171–199, 2007.
Christian Haefke, Marcus Sonntag, and Thijs Van Rens. Wage rigidity and job creation. Journal of mone-
tary economics, 60(8):887–899, 2013.
16
Marcus Hagedorn and Iourii Manovskii. The cyclical behavior of equilibrium unemployment and vacan-
cies revisited. American Economic Review, 98(4):1692–1706, 2008.
Marcus Hagedorn and Iourii Manovskii. Job selection and wages over the business cycle. American Eco-
nomic Review, 103(2):771–803, 2013.
Robert E Hall. Employment fluctuations with equilibrium wage stickiness. American economic review, 95
(1):50–65, 2005.
Robert E Hall and Paul R Milgrom. The limited influence of unemployment on the wage bargain. Ameri-
can economic review, 98(4):1653–74, 2008.
Hervé Le Bihan, Jérémi Montornès, and Thomas Heckel. Sticky wages: evidence from quarterly microe-
conomic data. American Economic Journal: Macroeconomics, 4(3):1–32, 2012.
Lars Ljungqvist and Thomas J Sargent. The fundamental surplus. American Economic Review, 107(9):
2630–65, 2017.
Pedro S Martins, Gary Solon, and Jonathan P Thomas. Measuring what employers do about entry wages
over the business cycle: A new approach. American Economic Journal: Macroeconomics, 4(4):36–55,
2012.
Espen R Moen. Competitive search equilibrium. Journal of political Economy, 105(2):385–411, 1997.
Fei Peng and W Stanley Siebert. Real wage cyclicality in italy. Labour, 22(4):569–591, 2008.
Barbara Petrongolo and Christopher A Pissarides. Looking into the black box: A survey of the matching
function. Journal of Economic literature, 39(2):390–431, 2001.
Christopher A Pissarides. The unemployment volatility puzzle: Is wage stickiness the answer? Economet-
rica, 77(5):1339–1369, 2009.
Aysegül Sahin, Joseph Song, Giorgio Topa, and Giovanni L Violante. Mismatch unemployment. American
Economic Review, 104(11):3529–64, 2014.
Robert Shimer. The cyclical behavior of equilibrium unemployment and vacancies. American economic
review, 95(1):25–49, 2005.
Donggyun Shin. Cyclicality of real wages among young men. Economics Letters, 46(2):137–142, 1994.
Frank Smets and Raf Wouters. An estimated dynamic stochastic general equilibrium model of the euro
area. Journal of the European economic association, 1(5):1123–1175, 2003.
Frank Smets and Rafael Wouters. Shocks and frictions in us business cycles: A bayesian dsge approach.
American economic review, 97(3):586–606, 2007.
17
A Additional Figures
Figure 6: Burning Glass Salaries Match OES Hourly Wages
22.
53
3.5
44.
5Lo
g M
edia
n H
ourly
Wag
e, O
ES
10 10.5 11 11.5 12Log Median Salary for Base Pay Annual Workers, Burning Glass
Notes: In both Burning Glass and the OES, the variable is the log of the median salary for hourly basepay workers, by 6-digit SOC cells. Burning Glass data is 2010-2016. The OES data is 2014-2016. The dataare binned into percentiles of the regressor, and weighted by employment shares in the OES at the 6-digitlevel. The regression slope, estimated from the underlying data, is 1.139.
18
Figure 7: Burning Glass Salaries Match QCEW Weekly Earnings
66.
57
7.5
Log
Ave
rage
Wee
kly
Ear
ning
s, Q
CE
W
10.6 10.8 11 11.2 11.4 11.6Log Median Salary for Base Pay Annual Workers, Burning Glass
Notes: In Burning Glass, the variable is the log of the median salary for hourly base pay workers, by CBSA.In the QCEW, the variable is the log of average weekly earnings, by CBSA. Burning Glass and QCEW dataare both 2010-2016. The data are binned into percentiles of the regressor, and weighted by employmentshares in the QCEW at the CBSA level. The regression slope, estimated from the underlying data, is 1.30.
19
Figure 8: Comparison of Employment Shares by Occupation, in Burning Glass and the OES
0 5 10 15Percent
Transportation and Material Moving
Sales and Related
Protective Service
Production
Personal Care and Service
Office and Administrative Support
Management
Life, Physical, and Social Science
Legal
Installation, Maintenance, and Repair
Healthcare Support
Healthcare Practitioners and Technical
Food Preparation and Serving Related
Farming, Fishing, and Forestry
Education, Training, and Library
Construction and Extraction
Computer and Mathematical
Community and Social Service
Business and Financial Operations
Building and Grounds Cleaning and Maintenance
Arts, Design, Entertainment, Sports, and Media
Architecture and Engineering OES Share
Burning Glass Share
Notes: In Burning Glass, the data is 2010-2016; in the OES, the data is 2014-2016. In both datasets, thecomparison is at the 2 digit SOC level, and excludes military.
20
B Additional Tables
Table 4: Comparison of OES and Burning Glass Wages, by 6-digit SOC Occupation
Dependent Variable: Log Median Hourly Wage by Occupation (OES)(1) (2) (3) (4)
Independent Variable:Log Median Salary 1.139*** 1.174*** 0.779*** 1.001***by Occupation (BG) (0.0945) (0.0678) (0.0883) (0.0899)BG Salary Type Base Pay, Annual Base Pay, Hourly Total Pay, Annual Total Pay, HourlyObservations 742 751 742 754
Notes: the dependent variable is the log median hourly wage, by 6-digit SOC occupation in the 2014-2016 Occupational Employment Statistics. The independent variable is the log median salary, by 6-digitSOC occupation in Burning Glass, for each salary type and pay frequency, for 2010-2016. The regressionis weighted least squares, weighted by 6-digit SOC occupation employment share in the OES. Robuststandard errors are in parentheses. One, two and three asterisks denote significance at the 10, 5 and 1percent levels, respectively.
21
Table 5: Comparison of QCEW and Burning Glass Wages, by CBSA
Dependent Variable: Log Average Weekly Earnings by CBSA (QCEW)(1) (2) (3) (4)
Independent Variable:Log Median Salary 1.295*** 1.390*** 1.069*** 0.900***by CBSA (BG) (0.0754) (0.127) (0.100) (0.149)BG Salary Type Base Pay, Annual Base Pay, Hourly Total Pay, Annual Total Pay, HourlyObservations 928 928 927 928
Notes: the dependent variable is average weekly earnings by CBSA, from the 2010-2016 QCEW. The inde-pendent variable is the median salary by CBSA, pay frequency and salary type, from the 2010-2016 Burn-ing Glass data. The regression is weighted least squares, weighted by CBSA employment in the QCEW.Robust standard errors are in parentheses. One, two and three asterisks denote significance at the 10, 5and 1 percent levels, respectively.
22
Table 6: Quarterly Posted Wage Cyclicality, Differenced by Job, Combined Statistical Area
Dependent Variable: Posted Wage Growth, by Job, CSA Level(1) (2) (3) (4) (5)
Independent Variable:Quarterly Unemployment Change, CSA -0.151 -0.360** -0.0558 0.151 -0.161
(0.165) (0.169) (0.301) (0.248) (0.275)Seasonal Dummies Y Y Y Y YDifference Length Dummies Y Y Y Y YTime Effects N Y N N NOES Weights N N Y N NCBSA Fixed Effects N N N Y NWinsorized N N N N YObservations 920086 920086 914062 919702 939725
Notes: the dependent variable is the log posted wage logwijt, for job i in CSA j at quarter t, from the 2010-2016 Burning Glass data. The independent variable is Ujt, the annual unemployment rate in CSA j atquarter t, from the 2010-2016 LAUS. We project Ujt onto log(Employmentjt), log CSA employment fromthe 2010-2016 QCEW. Posted wages are trimmed at the 1st and 99th percentile, except in column (3),where they are Winsorized at the 1st and 99th percentile. The controls are dummies for 6-digit SOC codeand 2-digit NAICS code. In column (4), the OES weights reweight the Burning Glass data to match the2014-2016 OES at the 6-digit SOC level. Standard errors are in parentheses, two-way clustered by CSA andquarter. One, two and three asterisks denote significance at the 10, 5 and 1 percent levels, respectively.
23
Table 7: Annual Posted Wage Cyclicality, Differenced by Job
Dependent Variable: Posted Wage Growth, by Job(1) (2) (3) (4) (5)
Independent Variable:Annual Unemployment Change -0.671 -0.700 -1.026 0.285 -0.516
(0.431) (0.451) (0.586) (0.581) (0.575)Difference Length Dummies Y Y Y Y YTime Effect N Y N N NOES Weights N N Y N NCBSA Fixed Effects N N N Y NWinsorized N N N N YNumber of Differenced Observations 496199 496199 492463 495532 506351
Notes: the dependent variable is percentage posted wage growth 100 × ∆ log (wijt) , for job i in CBSA jat year t, from the 2010-2016 Burning Glass data. Posted wages are averaged within each job-year. Theindependent variable is the change in Ujt, the annual unemployment rate in CBSA j at time t, from the2010-2016 LAUS. We project Ujt onto quarterly employment growth from the 2010-2016 QCEW. Postedwage growth is trimmed at the 1st and 99th percentile, except in column (5), in which they are Winsorizedat the 1st and 99th percentiles. In column (3), the OES weights reweight the Burning Glass data to matchthe 2014-2016 OES at the 6-digit SOC level. A job is an establishment by job title by pay frequency bysalary type unit. Standard errors are in parentheses, two-way clustered by CBSA and year. One, two andthree asterisks denote significance at the 10, 5 and 1 percent levels, respectively.
24
Table 8: Our Estimates of the Cyclicality of the Wage for New Hires Compared With the Literature
Unemployment Standard Data Source Standard FrequencySemi-elasticity Error Error Typeof New Hire Wage
Gertler et al (2016) -0.33 0.51 SIPP 1990-2012 Robust MonthlyHagedorn & Manovskii (2013) -1.78 0.50 NLSY 1979-2004 Robust QuarterlyHaefke et al (2013) -2.44 1.50 CPS 1984-2007 Robust QuarterlyBils (1985) -2.99 1.56 NLSY 1966-1981 Homoskedastic AnnualBarlevy (2001) -3.00 0.35 NLSY 1979-1993 Homoskedastic AnnualShin (1994) -3.80 1.14 NLSY 1966-1982 Homoskedastic AnnualOur Benchmark -0.09 0.09 BG 2010-2016 Clustered Quarterly
Notes: we adjust the estimates of Haefke, Sonntag & van Rens (2013) from the elasticity of wages withrespect to real labour productivity, to the semi-elasticity of wages with respect to unemployment, usingthe estimate of the sensitivity of unemployment to real labour productivity estimated by Pissarides (2009).We take the median estimate from each paper, and use the more negative value where there is ambiguity.We use the wage for new hires, and only consider workers transitioning out of unemployment where theseestimates are available. In Haefke et al (2013), the CPS data is from the Outgoing Rotation Group.
25